Skip to content

The open-source agent harness - the runtime layer that turns an LLM into a working agent.

max output token should fall back to db provided value if unset in AgentSpec (#220)

Srajan Asthanaa0cbdbe55681
diff --git a/packages/harness/src/agent-session/TurnResourceResolver.ts b/packages/harness/src/agent-session/TurnResourceResolver.ts
index 997dc845e..98ffc1d61 100644
--- a/packages/harness/src/agent-session/TurnResourceResolver.ts
+++ b/packages/harness/src/agent-session/TurnResourceResolver.ts
@@ -5,7 +5,7 @@ import type { ToolSource } from '../core/mcp/IMCPServer';
import { RemoteMCP, type RemoteMcpHeaders } from '../core/mcp/RemoteMCP';
import type { ToolSelectorConfig } from '../core/mcp/ToolSelectorPolicy';
import { ToolSet } from '../core/mcp/ToolSet';
-import type { AgentDefinition } from '../core/runtime/AgentDefinition';
+import type { AgentDefinition, ModelParams } from '../core/runtime/AgentDefinition';
import type { AgentInfo } from '../core/runtime/AgentThread.types';
import type { Sandbox, SandboxInfo } from '../core/sandbox/Sandbox';
import type { AgentTracing } from '../core/tracing/AgentTracing';
@@ -58,8 +58,11 @@ export class TurnResourceResolver<
constructor(
protected readonly deps: {
- /** Model name → client. Called once per resolved definition; may load provider config. */
- llm: (model: string) => Promise<ILLM>;
+ /** Model name → client and defaults. Called once per resolved definition; may load provider config. */
+ llm: (model: string) => Promise<{
+ modelClient: ILLM;
+ defaultModelParams: ModelParams;
+ }>;
/**
* MCP server name → connection details. Required to use spec.mcp_servers:
* the AgentSpec carries names only (no url/headers on the wire) — the
@@ -196,9 +199,10 @@ export class TurnResourceResolver<
// Sub-agents may request a different catalog model via agent_info.model;
// resolve that name so modelClient matches the override (not just a label).
const modelName = agentInfo?.model ?? spec.model.name;
+ const resolvedModel = await this.deps.llm(modelName);
return {
definition: {
- modelClient: await this.deps.llm(modelName),
+ modelClient: resolvedModel.modelClient,
// Sub-agents receive the delegated task as a user message; their system
// prompt is SUB_AGENT_IDENTITY (added by AgentThread), not user instructions.
instruction: agentInfo ? undefined : spec.instructions,
@@ -208,7 +212,10 @@ export class TurnResourceResolver<
role: 'user' as const,
content: m.content,
})),
- modelParams: spec.model.params,
+ modelParams: {
+ ...resolvedModel.defaultModelParams,
+ ...spec.model.params,
+ },
// Sub-agents should return free-form summaries to the parent, not the user-facing structured response.
responseFormat: agentInfo ? undefined : spec.response_format,
iterationLimit: spec.config.iteration_limit,
diff --git a/packages/harness/src/core/index.ts b/packages/harness/src/core/index.ts
index 45d460042..1b0c1a0e9 100644
--- a/packages/harness/src/core/index.ts
+++ b/packages/harness/src/core/index.ts
@@ -4,7 +4,7 @@
*/
// Runtime
-export type { AgentDefinition } from './runtime/AgentDefinition';
+export type { AgentDefinition, ModelParams } from './runtime/AgentDefinition';
export { AgentThread } from './runtime/AgentThread';
export type {
AgentInfo,
diff --git a/packages/harness/tests/agent-session/resolveAgentSpec.test.ts b/packages/harness/tests/agent-session/resolveAgentSpec.test.ts
index 7f0b6b49c..2ad56ad21 100644
--- a/packages/harness/tests/agent-session/resolveAgentSpec.test.ts
+++ b/packages/harness/tests/agent-session/resolveAgentSpec.test.ts
@@ -1,3 +1,4 @@
+import { AgentSpecSchema } from '../../src/agent-session/schemas/agentSpec';
import { EventType } from '../../src/agent-session/schemas/events';
import { Sessions } from '../../src/agent-session/Sessions';
import { InMemorySessionStore } from '../../src/agent-session/store/InMemorySessionStore';
@@ -7,7 +8,7 @@ import { makeAgentSpec, makeMockILLM, makeSilentLogger, makeTestResolver, mintTe
describe('TurnResourceResolver.resolveAgentSpec', () => {
it('fails closed when deps.agent is not wired for a named lookup', async () => {
const resolver = new TurnResourceResolver({
- llm: () => Promise.resolve(makeMockILLM()),
+ llm: () => Promise.resolve({ modelClient: makeMockILLM(), defaultModelParams: {} }),
mcp: () => Promise.reject(new Error('unused')),
mcpRequestTimeoutMs: 1_000,
mcpConnectTimeoutMs: 1_000,
@@ -18,6 +19,50 @@ describe('TurnResourceResolver.resolveAgentSpec', () => {
});
});
+describe('TurnResourceResolver.resolveAgentDefinition', () => {
+ it.each([
+ {
+ name: 'uses the resolved model default when the agent omits max_tokens',
+ resolvedModelParams: { max_tokens: 4096 },
+ agentModelParams: undefined,
+ expected: 4096,
+ },
+ {
+ name: 'lets the agent max_tokens override the resolved model default',
+ resolvedModelParams: { max_tokens: 4096 },
+ agentModelParams: { max_tokens: 8192 },
+ expected: 8192,
+ },
+ ])('$name', async ({ resolvedModelParams, agentModelParams, expected }) => {
+ const resolver = new TurnResourceResolver({
+ llm: () =>
+ Promise.resolve({
+ modelClient: makeMockILLM(),
+ defaultModelParams: resolvedModelParams,
+ }),
+ mcp: () => Promise.reject(new Error('unused')),
+ mcpRequestTimeoutMs: 1_000,
+ mcpConnectTimeoutMs: 1_000,
+ logger: makeSilentLogger(),
+ });
+ const spec = AgentSpecSchema.parse({
+ model: {
+ name: 'provider/model',
+ ...(agentModelParams === undefined ? {} : { params: agentModelParams }),
+ },
+ });
+
+ const { definition } = await resolver.resolveAgentDefinition({
+ spec,
+ thread_id: 'main',
+ signal: new AbortController().signal,
+ tracing: resolver.createTracing(),
+ });
+
+ expect(definition.modelParams?.['max_tokens']).toBe(expected);
+ });
+});
+
describe('SessionHandle.createTurn named resolve', () => {
it('loads the live AgentSpec through resolver.agent for ref sessions', async () => {
const store = new InMemorySessionStore();
diff --git a/packages/harness/tests/agent-session/testHelpers.ts b/packages/harness/tests/agent-session/testHelpers.ts
index 4a376473e..e9be2f4a6 100644
--- a/packages/harness/tests/agent-session/testHelpers.ts
+++ b/packages/harness/tests/agent-session/testHelpers.ts
@@ -90,7 +90,7 @@ export function makeTestResolver<TTurnCustom extends object = Record<string, nev
create: jest.fn().mockImplementation(() => emptyLlmStream(options?.usage)),
});
const base = new TurnResourceResolver<TTurnCustom>({
- llm: () => Promise.resolve(llm),
+ llm: () => Promise.resolve({ modelClient: llm, defaultModelParams: {} }),
mcp: name => {
return Promise.reject(new Error(`unexpected mcp lookup: ${name}`));
},
diff --git a/packages/harness/tests/agent-session/turnStream.test.ts b/packages/harness/tests/agent-session/turnStream.test.ts
index 2426cbc01..2c0c93d30 100644
--- a/packages/harness/tests/agent-session/turnStream.test.ts
+++ b/packages/harness/tests/agent-session/turnStream.test.ts
@@ -391,7 +391,11 @@ describe('TurnHandle.stream()', () => {
const closeSpy = jest.spyOn(sandbox, 'close').mockResolvedValue(undefined);
const logger = makeSilentLogger();
const resolver = new TurnResourceResolver({
- llm: () => Promise.resolve(makeMockILLM({ create: jest.fn().mockImplementation(() => emptyLlmStream()) })),
+ llm: () =>
+ Promise.resolve({
+ modelClient: makeMockILLM({ create: jest.fn().mockImplementation(() => emptyLlmStream()) }),
+ defaultModelParams: {},
+ }),
mcp: () => Promise.resolve({ url: 'http://localhost' }),
mcpRequestTimeoutMs: 60_000,
mcpConnectTimeoutMs: 5_000,
@@ -445,7 +449,7 @@ describe('TurnResourceResolver caches', () => {
expect(a).toBe(b);
}
})({
- llm: () => Promise.resolve(makeMockILLM()),
+ llm: () => Promise.resolve({ modelClient: makeMockILLM(), defaultModelParams: {} }),
mcp: () => Promise.resolve({ url: 'http://example.invalid' }),
mcpRequestTimeoutMs: 60_000,
mcpConnectTimeoutMs: 5_000,
@@ -461,7 +465,11 @@ describe('TurnResourceResolver caches', () => {
let sandboxCreates = 0;
const logger = makeSilentLogger();
const resolver = new TurnResourceResolver({
- llm: () => Promise.resolve(makeMockILLM({ create: jest.fn().mockImplementation(() => emptyLlmStream()) })),
+ llm: () =>
+ Promise.resolve({
+ modelClient: makeMockILLM({ create: jest.fn().mockImplementation(() => emptyLlmStream()) }),
+ defaultModelParams: {},
+ }),
mcp: () => Promise.resolve({ url: 'http://localhost' }),
mcpRequestTimeoutMs: 60_000,
mcpConnectTimeoutMs: 5_000,
diff --git a/packages/server/src/apis/turns.ts b/packages/server/src/apis/turns.ts
index 6716cbeda..cd9c42519 100644
--- a/packages/server/src/apis/turns.ts
+++ b/packages/server/src/apis/turns.ts
@@ -49,7 +49,7 @@ import { validateSandboxFilePath } from '../runtime/sandboxFilePath';
import {
buildTurnSandbox,
getMcpConnection,
- getModelProviderConfig,
+ getModelDetails,
resolveGitSkills,
resolveSandboxProvider,
} from '../runtime/sessionResources';
@@ -141,16 +141,19 @@ function createTurnResolver(deps: {
} = deps;
return new TurnResourceResolver({
llm: async name => {
- const providerConfig = await getModelProviderConfig({
+ const { providerConfig, defaultModelParams } = await getModelDetails({
tenant_id: TENANT_ID,
name,
store: modelProviderStore,
});
- return new VercelAILLM({
- providerConfig,
- logger,
- signal,
- });
+ return {
+ modelClient: new VercelAILLM({
+ providerConfig,
+ logger,
+ signal,
+ }),
+ defaultModelParams,
+ };
},
mcp: async name => {
const connection = await getMcpConnection({
diff --git a/packages/server/src/runtime/sessionResources.ts b/packages/server/src/runtime/sessionResources.ts
index c2acd6805..6696e0acb 100644
--- a/packages/server/src/runtime/sessionResources.ts
+++ b/packages/server/src/runtime/sessionResources.ts
@@ -9,6 +9,7 @@ import {
SkillMounter,
type AgentTracing,
type GitSkill,
+ type ModelParams,
type RemoteMcpHeaders,
type SandboxProvider,
type VercelAIProviderConfig,
@@ -43,10 +44,10 @@ export function parseModelFqn(name: string): { providerName: string; modelName:
}
/**
- * Load turn-ready LLM config for a configured FQN (`provider/model`).
+ * Load turn-ready model config and defaults for a configured FQN (`provider/model`).
* Malformed FQN or missing provider/model → HTTPException(422).
*/
-export async function getModelProviderConfig({
+export async function getModelDetails({
tenant_id,
name,
store,
@@ -54,7 +55,10 @@ export async function getModelProviderConfig({
tenant_id: string;
name: string;
store: IModelProviderStore;
-}): Promise<VercelAIProviderConfig> {
+}): Promise<{
+ providerConfig: VercelAIProviderConfig;
+ defaultModelParams: ModelParams;
+}> {
const parsed = parseModelFqn(name);
if (parsed === undefined) {
throw new HTTPException(422, {
@@ -77,14 +81,16 @@ export async function getModelProviderConfig({
// `buildLanguageModel` case fails to compile here.
const { type, base_url } = provider.manifest;
return {
- provider: type,
- name,
- modelId: model.model_id,
- baseUrl: base_url,
- // Custom providers may omit auth; adapters still require a string.
- apiKey: provider.manifest.auth?.api_key ?? '',
-
- headers: {},
+ providerConfig: {
+ provider: type,
+ name,
+ modelId: model.model_id,
+ baseUrl: base_url,
+ // Custom providers may omit auth; adapters still require a string.
+ apiKey: provider.manifest.auth?.api_key ?? '',
+ headers: {},
+ },
+ defaultModelParams: model.properties.max_output_tokens ? { max_tokens: model.properties.max_output_tokens } : {},
};
}
diff --git a/packages/server/tests/unit/runtime/sessionResources.test.ts b/packages/server/tests/unit/runtime/sessionResources.test.ts
index 4d97ae93f..f19541194 100644
--- a/packages/server/tests/unit/runtime/sessionResources.test.ts
+++ b/packages/server/tests/unit/runtime/sessionResources.test.ts
@@ -7,7 +7,7 @@ import { SqliteMcpServerStore } from '../../../src/db/sqlite/mcp-server-store/Sq
import { SqliteModelProviderStore } from '../../../src/db/sqlite/model-provider-store/SqliteModelProviderStore';
import { SqliteSandboxProviderStore } from '../../../src/db/sqlite/sandbox-provider-store/SqliteSandboxProviderStore';
import { SqliteSkillStore } from '../../../src/db/sqlite/skill-store/SqliteSkillStore';
-import { validateAgentSpec } from '../../../src/runtime/sessionResources';
+import { getModelDetails, validateAgentSpec } from '../../../src/runtime/sessionResources';
import type { ReasoningEffort } from '../../../src/schemas/modelProvider';
describe('validateAgentSpec', () => {
@@ -45,6 +45,20 @@ describe('validateAgentSpec', () => {
};
}
+ it('maps the configured model output limit to runtime max_tokens', async () => {
+ const stores = await setup();
+
+ await expect(
+ getModelDetails({
+ tenant_id: TENANT_ID,
+ name: 'test-provider/test-model',
+ store: stores.modelProviderStore,
+ }),
+ ).resolves.toMatchObject({
+ defaultModelParams: { max_tokens: 4096 },
+ });
+ });
+
it('rejects malformed model FQN with 422', async () => {
const stores = await setup();
await expect(